Robodebt and the cost of an unchecked automated decision.
How an Australian debt-raising scheme failed between 2015 and 2020, and what it teaches anyone who automates a decision about a person.
Terence KokExecutive Director, AI Governance & Assurance, Orion Five Engineering
About your speaker
Terence Kok
Executive Director, AI Governance & Assurance Practice, Orion Five Engineering
Former Chief AI and Innovation Officer, Meinhardt Group
Twenty-five years leading digital and AI transformation across Asia and the Middle East: national smart-city programmes, critical infrastructure and enterprise operations. Today I help leadership teams put AI into production with controls they can show a board, an auditor and a regulator.
BuiltA governed AI platform serving more than 6,000 engineers, with zero security incidents
CertifiedISO/IEC 42001 Lead Auditor Professional (BSI) · AIGP (SMU) · Registered Management Consultant
ServingCTCB CPSC Advisory Board · UN ESCAP AI for Developing Countries Forum
AwardsCDO Vision Award 2026 · Singapore's Most Influential Data & AI Leaders 2026 · Google Trailblazer Transformation Award 2024
Independent analysis from public sources. The views in this session are mine.
This session
Five years of automated debts, examined in eight parts.
How it beganA data-matching idea and a legal warning in 2014 and 2015
The mechanismHow income averaging produced debts that did not exist
Full scale20,000 notices a week from September 2016
Seven warningsEach one had an answer that let the scheme continue
The legal collapseFrom the Solicitor-General to the Federal Court
People and moneyThe human cost and the A$565 million net cost
The reckoningRoyal Commission, APSC, NACC and a second settlement
Lessons and controlsWhat boards and agencies can take from it
Case study · Robodebt
The short version
An automated scheme raised debts with no basis in law.
794,000
debts raised, about, against about 526,000 people
A$4.772bn
savings budgeted when the measure was funded
A$406m
savings achieved, about, at a total cost of A$971 million
A$565m
net cost to the Commonwealth, about, before two class action settlements
[1] Royal Commission report, pp. 401 to 402. Figures differ by source and definition; the Commission's are used throughout.
Case study · Robodebt
Why this case matters for AI
Robodebt used automation. The Commission says it did not use AI.
It was a rules-based system that matched data, applied averaging and generated debt notices with little or no human review.
Its lessons carry to AI because the failures were in how decisions about the system were made: the legal basis was never settled, the human check was designed out, the burden of proof moved to the public, and five years of warnings went unheeded.
[1] Royal Commission report, p. 472.
2014 to 2026
Twelve years from the first legal warning to the last finding.
Late 2014
Legal warning
Department lawyers say averaging may not fit the legislation.
2015
Funded
Cabinet is told no legislation is needed. Pilots run.
Scheme ends June 2020. Settlement approved June 2021.
Jul 2023
Royal Commission
Reports, with a sealed chapter of referrals.
2024 to 2026
Accountability
APSC, NACC and a second settlement.
[1] [2] [3] [4] [9] [24] Full chronology in the case study.
Case study · Robodebt
Part 1 · 2014 to 2015
How the scheme began.
The idea was to use data the government already held. What was new was letting an annual figure stand in for fortnightly ones.
The data idea
Two government records that measure income differently.
[1] Royal Commission report.
Case study · Robodebt
Late 2014 to March 2015
The legal caveat was lost on the way to Cabinet.
Late 2014Department of Social Services lawyers advise that averaging "might not be consistent with the legislative framework".
February 2015The Social Services Minister, Scott Morrison, signs an executive minute noting that legislative change would be needed.
March 2015The New Policy Proposal to Cabinet says legislation is not required and does not mention averaging.
[1] Royal Commission report. The Commission called the statement false and found Mr Morrison allowed Cabinet to be misled because he did not make the obvious inquiry.
2015 pilot
The pilot gave a warning sign that nobody evaluated.
40%
About 40 per cent of people contacted in the 2015 pilot made contact back.
What followed
The pilot was not properly evaluated.
The online screens were never tested with users.
The measure was funded in the 2015 to 2016 Budget as strengthening the integrity of welfare payments, and extended in later budget updates.
[1] Royal Commission report.
Case study · Robodebt
Part 2 · The mechanism
How averaging produced debts.
Entitlement depends on what a person earned in each fortnight. Averaging assumed the same amount every fortnight.
One year, twenty-six fortnights
Averaging put income into fortnights when none was earned.
What was earned: $18,000 in ten fortnights, nothing owedWhat the scheme assumed in the fortnights on benefit
Illustrative figures. The mechanism follows the Royal Commission's account of averaging ATO annual income across fortnights [1] [27].
Case study · Robodebt
Five changes from the manual process
A lead for an officer became a notice generated by the system.
Before
Under the online scheme
ATO annual income data was a lead for an officer to check.
Unconfirmed ATO annual (PAYG) data became the primary evidence of income.
The department had to establish any debt.
The person had to disprove it, often with payslips several years old.
Entitlement was assessed on actual fortnightly income, as the law requires.
Annual income was averaged across fortnights automatically where the person did not supply records.
Compliance officers handled each case.
The process was designed to run online "to reduce (and preferably obviate) human involvement".
Reviews looked at recent periods.
Recovery reached back about five years.
[1] Royal Commission report, pp. xxiii to xxiv.
Case study · Robodebt
The process maps
When a person did not respond, every later step ran automatically.
01Data match
ATO annual income compared with fortnightly reports
02Silence as consent
No response was treated as accepting the ATO figure
03Entitlement
Recalculated automatically, with averaging
04Debt and penalty
Debt plus a 10 per cent penalty, calculated automatically
05Notice
Generated automatically, with no compliance phone number
The letters left out a phone number by design, to push people online. The Commission described the penalty as a "euphemistically called" recovery fee and found no evidence anyone had considered its legal basis.
[1] Royal Commission report, process maps put before the Commission.
Case study · Robodebt
The burden of proof
People had to disprove the machine with records they had been told to discard.
What was askedPayslips going back years, to rebut a figure the system had estimated.
What the department had saidIts own website told people to keep payslips for six months.
What it didTurned an estimate into a verdict that the person had to overturn.
[1] Royal Commission report.
Part 3 · 2016 to 2017
The scheme at full scale.
The Online Compliance Intervention began in July 2016 and was fully rolled out in September.
Volume
The online scheme issued in a week what the manual process handled in a year.
Manual processabout 20,000 a year, so about 385 a week
385
Online schemenotices a week, from Sep 2016
20,000
Per week, the online scheme ran at about 52 times the manual rate. In November 2016 the Minister for Human Services announced that new technology was helping raise $4.5 million in welfare debts a day.
[1] Royal Commission report. The weekly manual rate (20,000 divided by 52) is arithmetic on the Commission's figures.
Case study · Robodebt
December 2016 to January 2017
Within six months, three in four people in the online phase had averaged debts.
76%
of people in the online phase had debts raised from averaged earnings by 27 January 2017
99,404
people that 76 per cent represents
A$11.6m
paid to private debt collectors, about, between 2015 and 2021
By December the news was full of people with debts they did not owe, and the Australian Council of Social Service wrote to the Minister. Debts went to collectors without checking that the person had received the notice.
[1] Royal Commission report. [27] The Conversation, January 2017.
Case study · Robodebt
Part 4 · 2014 to 2019
Seven warnings.
The scheme ran for three more years after the first public alarm. Each warning had an answer that let it continue.
Warning and response
Every warning was answered in a way that kept the scheme running.
When
Warning
Response
Late 2014
Department lawyers advise averaging may not be consistent with the legislation.
The March 2015 Cabinet proposal says no legislation is needed.
Dec 2016
Media saturated with stories of wrong debts; ACOSS writes to the Minister.
The scheme continues at 20,000 notices a week.
Apr 2017
Ombudsman reports on the system, without examining whether averaging is lawful.
Ministers cite the report as evidence the scheme is lawful.
Jun 2017
Senate committee recommends putting the scheme on hold.
The scheme continues.
2017
About 29 tribunal decisions reject averaging by the end of May.
Individual cases fixed; the decisions are not acted on more widely.
Aug 2018
Draft external legal advice says averaging is not permissible.
Never finalised or acted on.
Nov 2019
Commonwealth concedes in Federal Court that an averaged debt was not validly made.
Averaging-only debts stop. The scheme closes in June 2020.
[1] [2] [14] [18]
Case study · Robodebt
Warning three · April 2017
The oversight body was given a partial picture and then used as cover.
What the Ombudsman received
A 2017 internal view that averaging could be used "as a last resort", citing no law
Not the 2014 legal advice
The Commission found
Officers' explanation to the Ombudsman was untrue.
How the report was used
It did not examine legality
Ministers then cited it as evidence the scheme was lawful
[1] Royal Commission report. [14] Commonwealth Ombudsman, 10 April 2017.
Case study · Robodebt
Warning five · 2017
Twenty-nine tribunal losses were handled one case at a time.
29 decisions rejecting averaging by the end of May 2017. Each was fixed for that one person.
By the end of May 2017 about 29 first-tier tribunal decisions had rejected averaging. The department fixed each appealed case and otherwise carried on. An error report that arrives one case at a time can be absorbed indefinitely unless someone is required to read the pattern.
[1] Royal Commission report. [2] Prygodicz (No 2) [2021] FCA 634.
Case study · Robodebt
Part 5 · 2019 to 2021
The legal collapse.
The legal position, known inside government since 2014, came out in court.
Seven steps in two years
Once the law was tested, the scheme fell within months.
Mar 2019
Government Solicitor
Averaging cannot establish a debt.
Sep 2019
Solicitor-General
Averaged ATO data alone cannot support whether a debt exists or its amount.
19 Nov 2019
Debts stop
No more debts on averaged data alone. A class action is filed the same day.
27 Nov 2019
Amato
The Commonwealth consents: the averaged debt was not validly made.
30 Jun 2020
Scheme closes
Refunds begin in July.
Nov 2020
Settlement
The class action settles on the day set for trial.
Jun 2021
Approved
The Federal Court records the Commonwealth's admission.
[1] Royal Commission report. [2] Prygodicz (No 2) [2021] FCA 634.
Case study · Robodebt
Federal Court, June 2021
The Commonwealth "did not have a proper legal basis to raise, demand or recover" debts based on averaging.
The Commonwealth's admission, as recorded by Justice Murphy in approving the class action settlement. The settlement added A$112 million.
[2] Prygodicz v Commonwealth of Australia (No 2) [2021] FCA 634.
Case study · Robodebt
Part 6 · People and money
The human cost.
The Commission heard from people who received debts they did not owe and could not disprove, and from families of people who had died.
What the Commission recorded
The record includes deaths, and a minister who knew.
The Commission records that Rhys Cauzzo, aged 28, died in January 2017 after debts raised by averaging in the manual phase that preceded the online scheme had been referred to a debt collector.
It records that Jarrad Madgwick, aged 22, died in May 2019, and that his mother identified a debt notice as a precipitating factor. The Commission notes his debt does not appear to have come from averaging.
The Commission found that by July 2017 the Minister for Human Services, Alan Tudge, knew that at least two people had died by suicide, and that no comprehensive review followed.
For the roomHarm indicators were never reported beside the savings figures.
[1] Royal Commission report. If this topic affects you, support is available: in Australia, Lifeline 13 11 14; in Singapore, SOS 1767; elsewhere, findahelpline.com.
Case study · Robodebt
The Commission's summary
"Empathy could not be programmed into the Scheme."
Royal Commission into the Robodebt Scheme, 2023
[1] Royal Commission report.
Case study · Robodebt
The money · A$ millions
A projected saving of A$4.8 billion ended as a net cost.
4,772
Budgeted savingswhen funded
406
Savings achievedabout
971
Total costof running the scheme
565
Net costaccepted by the Commonwealth
[1] Royal Commission report, p. 401. Net cost is before the two class action settlements.
Case study · Robodebt
Refunds, write-offs and settlements · A$ millions
A$1.751 billion was written off, and the refunds are part of it.
Written off in total
Refunded, A$746m to about 381,000 people
1,751
Settlement, 2021class action
112
Settlement, 2026A$475m compensation plus costs
548.5
[1] pp. xxviii and 402 (the written-off total includes the refunds; never add them). [2] [24] [25] [26]. Bars share one scale.
Case study · Robodebt
Part 7 · 2022 to 2026
The reckoning.
A Royal Commission, led by Catherine Holmes AC SC, was established in August 2022 and reported on 7 July 2023.
Royal Commission, July 2023
The Commission's language was unusually direct.
"A crude and cruel mechanism, neither fair nor legal."Report, p. xxix
"A costly failure of public administration, in both human and economic terms."Report
"The automation used in the Scheme at its outset, removing the human element, was a key factor in the harm it did."Report, p. 488
"A myriad of ways that things can go wrong through venality, incompetence and cowardice."Report, closing
[1] Royal Commission report. A sealed chapter referred individuals to the APSC, the NACC, the Australian Federal Police and a legal regulator.
Case study · Robodebt
Accountability
Six outcomes over five years.
Body
Outcome
Federal Court, 2021
Settlement approved. The Commonwealth admitted it had no proper legal basis to raise, demand or recover averaged debts.
Royal Commission, 2023
Adverse findings about ministers and officials; a sealed chapter of referrals to four bodies.
APS Commission, 2024
16 inquiries. 12 people found in breach of the Code of Conduct. Four still employed were sanctioned, from reprimands and fines to demotion; the rest had left the service.
NACC, 2024
Declined to investigate, saying new evidence was unlikely. The Inspector later found officer misconduct in that decision.
NACC, 2026
After independent reconsideration: two former senior officials engaged in serious corrupt conduct. No referral for prosecution: the decisive evidence was compelled and could not be used in court.
Federal Court, 2026
Second settlement approved 23 June: A$475 million in compensation, about A$548.5 million with costs, for about 125,000 registered people.
[1] [2] [4] [5] [6] [9] [10] [24] [25] [29] [30]
Case study · Robodebt
APS Commission, September 2024
Twelve of sixteen public servants breached the Code of Conduct.
4in breach and still employed: sanctioned, from reprimands and fines to demotion
8in breach who had already left the service
4not found in breach
[4] Australian Public Service Commission, outcomes of 13 September 2024. [28] ACS Information Age. Individuals are not named.
Case study · Robodebt
National Anti-Corruption Commission
The anti-corruption finding took two attempts.
Jun 2024
Declines
Decides not to investigate the referrals.
Oct 2024
Inspector
Finds officer misconduct in that decision.
Dec 2024
Delegate
An independent delegate is appointed.
Feb 2025
Reopens
Decides the referrals should be investigated.
Mar 2026
Finds
Two former senior officials engaged in serious corrupt conduct.
Mar 2026
Clears four
Four others, including Mr Morrison, did not act corruptly.
One official misled officers preparing the 2015 Cabinet submission; one misled the Ombudsman in 2017. No referral for prosecution. The Royal Commission's finding about Mr Morrison and the NACC finding apply different legal tests, and both stand.
[5] [6] [7] [8] [9] [10]
Case study · Robodebt
Part 8 · Automated decisions
What the Royal Commission recommended.
The government accepted all 56 recommendations in November 2023, in full or in principle. Two deal directly with automation.
Recommendations 17.1 and 17.2
A legal framework for automation and a body to audit it.
17.1
A consistent legal framework
A clear path to review
Plain-language notice on agency websites that automated decision-making is used
Business rules and algorithms "made available, to enable independent expert scrutiny"
17.2
A body to monitor and audit
Technical soundness
Fairness
Bias
Usability
[1] Royal Commission report, pp. 486 to 488. [3] Government response, November 2023. The report's text refers to 57 recommendations; its numbered list and the government's response contain 56.
Case study · Robodebt
Progress since 2023
Three years on, the framework has not been legislated.
Nov 2024
Consultation
Attorney-General's Department consults on a framework to January 2025.
Mar 2025
Ombudsman guide
Automated Decision-Making Better Practice Guide updated.
Aug 2025
Summary
Consultation summary published.
Jul 2026
Priority
Government asks the Attorney-General to lead an ADM framework.
7 Sep 2026
Private bill
Still before the House. No government bill as at 12 October 2026.
10 Dec 2026
Privacy Act
Duties to disclose certain automated decisions start.
[16] [20] [21] [22] [23] [31] [32]
Case study · Robodebt
For AI
Every failure in Robodebt has a counterpart in AI deployment.
A rules engine that averages income and a model that predicts it both produce an estimate that someone has to decide whether to act on. Four questions apply to any AI system that decides or recommends something about a person.
01
Is there a legal basis for acting on this output?
02
Who checks it before it reaches the person?
03
Who carries the burden of correcting it?
04
Who reads the pattern of complaints and appeals?
Case study · Robodebt
My reading
Two features make AI riskier on the same points.
01
Harder to explain
A probabilistic model is harder to explain than a formula, so the "business rules made available" that the Commission asked for are harder to produce.
02
Faster and larger
The speed and scale that made the online scheme attractive are greater with AI. An error reaches more people before anyone reads the pattern.
SoThe Commission's recommendations were written for automation. They read as a minimum for AI.
Case study · Robodebt
Lessons
What boards and agencies can take from it.
Eight lessons, each tied to a finding. The lessons are my reading of those findings.
Lessons 1 and 2
Settle the law first, and keep a person where an error reaches a person.
01
Settle the legal basis before you automate
The lawyers' caveat was in the 2015 minute and missing from the Cabinet proposal.
Do thisGet written confirmation that the method is lawful, and make it travel with the funding decision.
02
Keep a human check where an error reaches a person
Removing the human element was a key factor in the harm.
Do thisDecide which outputs a person reviews before they take effect, and fund reviewers when volume rises.
Case study · Robodebt
Lessons 3 and 4
Leave the burden of proof with the organisation, and read appeals as a pattern.
03
Do not move the burden of proof to the public
Asking people to disprove a machine's figure with records they were told to discard turns an estimate into a verdict.
Do thisThe organisation establishes the decision; the person is never asked to disprove an estimate.
04
Read appeals and complaints as a pattern
Twenty-nine tribunal decisions were handled one at a time.
Do thisA named senior owner reports adverse external decisions in aggregate, with the error rate they imply.
Case study · Robodebt
Lessons 5 and 6
Protect legal advice, and count harm beside savings.
05
Protect the independence of legal advice
Advice that averaging was unlawful existed in 2014, 2018 and 2019.
Do thisLegal advice on an automated program goes to the board or minister directly, not through the team that owns the program.
06
Measure harm as well as savings
The scheme was judged on dollars recovered while deaths, complaints and tribunal losses went uncounted.
Do thisPut harm indicators beside the financial ones from the first month.
Case study · Robodebt
Lessons 7 and 8
Publish the rules, and know what a clean review covered.
07
Publish the rules and give reasons
The notices gave no reasons. The Commission asks for business rules and algorithms to be open to expert scrutiny.
Do thisFor AI, document what the system does and does not consider, and explain each decision to the person affected.
08
Check what an oversight body was asked
The Ombudsman's report was used as proof of legality on a question it had not examined.
Do thisTreat a clean review as covering only what the reviewer was asked to look at.
Case study · Robodebt
Controls checklist
Eight controls a board can ask to see before an automated decision goes live.
#
Control
Evidence to ask for
1
Legal basis confirmed
Written legal advice on the method, attached to the funding paper
2
Human review point
The outputs a person must review, and the reviewer capacity at peak volume
3
Burden of proof
Confirmation that no person is asked to disprove an estimate
4
Adverse-decision register
Appeals, complaints and reversals reported in aggregate, with an error rate
5
Independent legal channel
Legal advice that reaches the board or minister directly
6
Harm indicators
Harm measures reported beside savings from the first month
7
Published rules and reasons
Business rules on record; a reason given with every decision
8
Scope of any review
The questions each oversight body was asked, and what it was shown
My call, drawn from the findings in [1]. Free tools: the agent risk assessment matrix and the oversight ceiling calculator at terencekok.com/resources.
Case study · Robodebt
For discussion
Four questions to take back to your own organisation.
01
Which automated or AI output in your organisation reaches a person with no one checking it first?
02
Where would a pattern of appeals or complaints about that output be reported, and to whom?
03
Who last confirmed in writing that the method behind it is lawful?
04
If the savings target and the harm indicators disagreed, which one would your board see first?
Case study · Robodebt
Evidence and methodology
Each claim is graded by what it rests on.
Claim
Source
Grade
"A crude and cruel mechanism, neither fair nor legal"
Royal Commission report, p. xxix
Finding of a Royal Commission
The Scheme did not use AI; it used automation
Royal Commission report, p. 472
Finding of a Royal Commission
A$746m refunded; A$1.751bn written off in total
Report, pp. xxviii and 402. The total includes the refunds.
Measured
A$4.772bn budgeted; about A$406m achieved; net cost about A$565m
Report, p. 401
Measured
About 794,000 debts against about 526,000 people
Report, p. 402. The 2021 judgment's class is narrower.
Measured; differs by definition
A$475m second settlement, approved 23 June 2026
Gordon Legal and AAP; approval judgment not read
Reported
The averaging chart
Mechanism per the Commission; figures invented
Illustrative
Lessons and controls
My reading of the cited findings
My call
Case study · Robodebt
Sources 1 of 2
Sources 1 to 16.
1Royal Commission into the Robodebt Scheme (C. Holmes AC SC). (2023, 7 July). Report. robodebt.royalcommission.gov.au
2Prygodicz v Commonwealth of Australia (No 2) [2021] FCA 634 (Murphy J, 11 June 2021). Federal Court of Australia. judgments.fedcourt.gov.au
3Australian Government. (2023, November). Government response to the Royal Commission into the Robodebt Scheme. pmc.gov.au
4Australian Public Service Commission. (2023 to 2024). Robodebt Code of Conduct process (updates of 3 August 2023 to 13 September 2024). apsc.gov.au
5National Anti-Corruption Commission. (2024, 6 June). NACC decides not to pursue Robodebt Royal Commission referrals [Media release]. nacc.gov.au
6National Anti-Corruption Commission. (2024, 30 October). Public statement: NACC to reconsider decision not to investigate Robodebt referrals. nacc.gov.au
7National Anti-Corruption Commission. (2024, 13 December). Independent delegate appointed to reconsider Robodebt referrals. nacc.gov.au
16Commonwealth Ombudsman. (2025, March). Automated decision-making better practice guide. ombudsman.gov.au
Sources read on 11 October 2026, time-sensitive items re-checked 12 October 2026; some Australian Government documents through Internet Archive captures. Full list with links: terencekok.com/research/case-studies/robodebt/
Case study · Robodebt
Sources 2 of 2
Sources 17 to 32.
17Senate Community Affairs References Committee. (2017). Design, scope, cost-benefit analysis, contracts awarded and implementation associated with the Better Management of the Social Welfare System initiative [Inquiry page]. aph.gov.au
18Senate Community Affairs References Committee. (2017, 21 June). Report: List of recommendations. aph.gov.au
19Senate Community Affairs References Committee. (2022, May). Accountability and justice: Why we need a Royal Commission into Robodebt. aph.gov.au
20Attorney-General’s Department. (2024 to 2025). Automated decision-making reform [Consultation]. consultations.ag.gov.au
21IDM. (2026, 11 September). Bill would ban agencies from automating decisions. idm.net.au
22Dawson, S., Lau, E., & Cowan-Dillon, L. (2025, 14 January). Practical implications of the new transparency requirements for automated decision making. Jackson Wells Samuel (JWS). jws.com.au
23Huggins, A. (2025, 24 November). Regulating automated government decision-making: An Australian perspective. AUSPUBLAW. auspublaw.org
24Gordon Legal. (2026). Robodebt class action appeal settlement (FAQ last updated 2 October 2026; accessed 12 October 2026). gordonlegal.com.au
25Seeto, T. (2026, 24 June). Centrelink recipients win “historic” $475 million payout as Robodebt class action settlement approved. Yahoo Finance Australia. au.finance.yahoo.com
26Grattan, M. (2025, 4 September). Government settles Robodebt class action appeal for $475 million in compensation. The Conversation. theconversation.com
27Whiteford, P. (2017, 11 January). Note to Centrelink: Australian workers’ lives have changed. The Conversation. theconversation.com
28Sadler, D. (2024, 16 September). “Failure of government”: Public servants breached code over Robodebt. ACS Information Age. ia.acs.org.au
29Garbutt-Young, L. (2026, 23 June). Australia’s biggest class action settlement approved. Australian Associated Press, in The Canberra Times. canberratimes.com.au
30Lewis, A. (2026, 19 March). The “catch-22” protecting Robodebt bureaucrats from criminal prosecution. ABC News. abc.net.au
31Colley, A. (2026, 20 July). Feds move to regulate automated decision-making. iTnews. itnews.com.au
32Parliament of Australia. (2026). Automated Decision-Making (Safeguards and Transparency) Bill 2026 [Bill homepage; status “Before Reps”, checked 12 October 2026]. aph.gov.au
Sources read on 11 October 2026, time-sensitive items re-checked 12 October 2026; some Australian Government documents through Internet Archive captures. Full list with links: terencekok.com/research/case-studies/robodebt/
Case study · Robodebt
Case study · Robodebt
Settle the law, keep the human check, and read the pattern.
